Arnold Peter
Founding Partner at Peter Law Group
Date: October 7, 2026
Originally published on LinkedIn
Summary: Arnold Peter uses Steven Spielberg’s Minority Report to examine a difficult question: if artificial intelligence becomes more accurate than human judges and juries, should it be allowed to decide guilt? The article acknowledges human error and bias while arguing that statistical accuracy cannot replace reasonable doubt, dissent, accountability and mercy. Through a fictional AI investigation, Peter explores why powerful analytical tools should assist criminal justice while leaving the responsibility for a verdict with human beings.
What if artificial intelligence becomes better at determining guilt than we are? Not merely faster. Not simply better at searching databases, comparing fingerprints or identifying inconsistencies in testimony. Imagine an artificial intelligence capable of reviewing every photograph, DNA result, police report, witness statement, telephone record, surveillance video and expert opinion in a criminal prosecution—and then reaching the correct conclusion more often than a judge or jury. Would we let it decide who goes to prison?
The instinctive answer is no. But that answer becomes considerably more difficult when we confront an uncomfortable truth: human beings are not particularly reliable judges either. We forget. We misidentify people. We bring prejudice into the courtroom. We become emotional. We are influenced by confidence, appearance and storytelling. Judges make mistakes. Prosecutors make mistakes. Jurors make mistakes. Empirical studies of sitting trial judges have found that even experienced jurists are susceptible to cognitive illusions such as anchoring and hindsight bias. (Chris Guthrie, Jeffrey J. Rachlinski & Andrew J. Wistrich, Inside the Judicial Mind, 86 Cornell L. Rev. 777 (2001).) Related research has found that trial judges, like other people, harbor implicit racial bias. (Jeffrey J. Rachlinski et al., Does Unconscious Racial Bias Affect Trial Judges?, 84 Notre Dame L. Rev. 1195 (2009).) Innocent people have spent decades in prison because human beings sincerely believed they were guilty. The National Registry of Exonerations has documented thousands of exonerations in the United States since 1989, and one peer-reviewed study conservatively estimated that at least 4.1 percent of defendants sentenced to death would be exonerated if they remained under sentence of death indefinitely. (Samuel R. Gross et al., Rate of False Conviction of Criminal Defendants Who Are Sentenced to Death, 111 Proc. Nat’l Acad. Sci. 7230 (2014).)
That is what makes Steven Spielberg’s 2002 film Minority Report so relevant to the age of artificial intelligence. The movie is usually remembered as a warning against predicting crime before it happens. Its more interesting warning, however, may be about something else entirely: what happens when we create a system so accurate that we stop believing the dissenting voice? That question is no longer science fiction.
Minority Report’s Most Frightening Idea Wasn’t Predicting Crime
In Minority Report, Tom Cruise plays John Anderton, an officer in a futuristic “Precrime” division that arrests people before they commit murder. The system depends upon three human “precogs” whose visions supposedly reveal future crimes with extraordinary accuracy.
Normally, the three agree. Occasionally, however, one sees a different future. That conflicting prediction is the minority report, and because the system depends upon society believing in its near infallibility, the disagreement becomes profoundly inconvenient.
Lawyers should immediately recognize the problem because American jurisprudence is built around the opposite proposition: disagreement matters. Appellate courts publish majority opinions, concurrences and dissents. Juries deliberate because different human beings can reasonably interpret the same evidence differently. Since the Supreme Court’s decision in Ramos v. Louisiana, 590 U.S. 83 (2020), the Constitution requires unanimous verdicts for serious criminal convictions in both federal and state courts.
That means one juror can prevent eleven others from convicting someone. What Minority Report treats as an anomaly to be suppressed, our criminal justice system can treat as reasonable doubt. The question AI presents is whether we will continue to value that dissent once the machine appears statistically more reliable than the dissenter.
Before We Defend Human Judgement, We Should Admit How Bad It Can Be
There is a temptation in discussions about AI to romanticize human decision-making. We should resist it because the history of wrongful convictions provides devastating evidence of just how unreliable people can be.
Eyewitness identification is perhaps the clearest example. According to the Innocence Project, eyewitness misidentification was involved in 69 percent of the 375 DNA exonerations it studied from 1989 through 2020. Those innocent individuals served an average of approximately 14 years before exoneration, and 21 had spent time on death row. (Innocence Project, DNA Exonerations in the United States (1989-2020).) Professor Brandon Garrett’s study of the first 250 DNA exonerations found eyewitness misidentification in 76 percent of those cases. (Brandon L. Garrett, Convicting the Innocent: Where Criminal Prosecutions Go Wrong (Harvard Univ. Press 2011).) The National Academy of Sciences reached a similar conclusion, cautioning that eyewitness memory is malleable and vulnerable to error. (National Research Council, Identifying the Culprit: Assessing Eyewitness Identification (2014).)
Those numbers expose something deeply unsettling about human judgment. The witnesses responsible for wrongful identifications were not necessarily lying. Many sincerely believed they had identified the person who committed the crime. They could look across a courtroom, point toward a defendant and confidently tell twelve jurors: That is the person I saw. And they could be completely wrong.
Human memory is not a security camera. Stress, fear, distance, lighting, trauma, the passage of time and suggestive identification procedures can influence what someone remembers. (Elizabeth F. Loftus, Eyewitness Testimony (Harvard Univ. Press 1979).) Cross-racial identifications are especially prone to error. (Christian A. Meissner & John C. Brigham, Thirty Years of Investigating the Own-Race Bias in Memory for Faces: A Meta-Analytic Review, 7 Psychol. Pub. Pol’y & L. 3 (2001).) Worse, confidence and accuracy are not necessarily the same thing, yet jurors naturally find a confident witness persuasive. Research indicates that a witness’s confidence at an initial, properly conducted identification can be informative, but confidence expressed later at trial is often inflated and far less reliable. (John T. Wixted & Gary L. Wells, The Relationship Between Eyewitness Confidence and Identification Accuracy: A New Synthesis, 18 Psychol. Sci. Pub. Int. 10 (2017).)
AI therefore enters criminal justice not as an alternative to some pristine system of human rationality. It enters a system already populated by imperfect witnesses, imperfect police officers, imperfect lawyers, imperfect judges and imperfect jurors, all of whom are susceptible to the tunnel vision and confirmation bias that scholars have identified as recurring causes of wrongful convictions. (Keith A. Findley & Michael S. Scott, The Multiple Dimensions of Tunnel Vision in Criminal Cases, 2006 Wis. L. Rev. 291.) That changes the question dramatically.
Here Is the Uncomfortable Question: What If the Machine Really Is Better?
Suppose we eventually develop an AI system that makes fewer mistakes than human beings. Imagine a system that simultaneously analyzes DNA probabilities, cell-site information, photographs, witness statements, vehicle movements, financial transactions, surveillance footage and millions of other data points. It identifies inconsistencies humans missed and recognizes patterns no detective could realistically discover.
In my article, The Verdict Must Remain Human, I explored that possibility through a fictional future Supreme Court case, United States v. Mercer. The hypothetical involves an FBI system called FORECAST that reexamines a decades-old murder by analyzing DNA probabilities, location records, witness statements, vehicle movements, communications and patterns from thousands of solved homicides. It ultimately identifies Daniel Mercer as the likely perpetrator with a confidence score of 96.7 percent.
That hypothetical deliberately makes the easy answer difficult. FORECAST is not a cartoon villain. It is useful. It may even be extraordinarily accurate. Investigators follow its leads and uncover additional conventional evidence.
Now change the hypothetical slightly. Suppose rigorous studies establish that FORECAST reaches the correct result 99 percent of the time while human juries reach it 95 percent of the time. If replacing juries with AI would result in fewer innocent people being convicted, would insisting upon human judgment actually become unethical? That is the question we should be willing to confront rather than avoid.
The Majority Can Be Wrong—and Sometimes One Person Sees What Everyone Else Missed
The title Minority Report captures something fundamental about justice. A system can be overwhelmingly accurate in the aggregate and catastrophically wrong about one individual.
The law understands that possibility. Eleven jurors may look at the evidence and see guilt while the twelfth sees something everyone else has overlooked. Perhaps the timeline makes no sense. Perhaps an eyewitness identification seems too convenient. Perhaps the prosecution’s theory requires an assumption that has never actually been proved.
That twelfth juror might be unreasonable, stubborn or simply wrong. But the juror might also be the only person in the room who is right.
Our system tolerates that inefficiency because criminal justice does not merely ask which conclusion is statistically most probable. The prosecution carries the burden of proving guilt beyond a reasonable doubt, and unanimity forces the majority to confront the minority rather than simply outvote it. That is not computational efficiency. It is institutional humility.
Artificial intelligence could threaten that humility precisely because of its accuracy. A judge may find it relatively easy to disagree with another judge. A juror may disagree with another juror. But imagine disagreeing with an AI system that has analyzed ten million comparable cases and reports a 99.4 percent probability of guilt. The better AI becomes, the harder it may become for a human being to say: I still think the machine is wrong.
AI May Eliminate Human Bias—or Give It a More Impressive Vocabulary
There is another reason to resist assuming that machines will rescue us from ourselves. Artificial intelligence learns from information human beings created.
Historical criminal justice data reflect decisions about which neighborhoods police patrolled, whom officers stopped, whom prosecutors charged, which defendants received favorable plea agreements, which crimes were aggressively investigated and which witnesses were believed. If those decisions contained bias, the historical record can contain it as well.
AI can therefore reproduce human prejudice while making it look less human. A biased detective may sound biased. A biased algorithm may produce a number to three decimal places.
That appearance of mathematical neutrality is especially dangerous because numbers carry authority. My earlier work on AI and criminal justice has emphasized precisely this problem: a prejudiced human judgment may reveal itself as prejudice, while a prejudiced probability score can masquerade as science.
AI can undoubtedly reduce some forms of human bias. But it can also industrialize them. A biased police officer can distort one investigation at a time. A biased model can reproduce the same distortion across thousands of cases before anyone realizes what happened.
The Machine Can Calculate Guilt. But Can It Understand Justice?
Even if those technological problems could be solved, something deeper would remain. Justice is not simply an exercise in determining statistical probability. It requires judgment about individual human beings, and human judgment contains qualities that cannot easily be separated into “good” and “bad.”
We bring emotion into the courtroom. We bring cultural assumptions, personal experiences, prejudices, fears and sympathies. Those characteristics can distort justice and sometimes produce terrible results. But from the same imperfect humanity come skepticism, courage, empathy, compassion, moral intuition, forgiveness and mercy.
Our legal system does not regard those qualities as irrelevant defects to be engineered away. Judges consider mitigation. Prosecutors exercise discretion. Governors grant clemency. Juries assess credibility. Sentencing courts consider circumstances that distinguish one human being from another. The law recognizes rehabilitation and the possibility that people can change.
An AI system may eventually calculate recidivism risk more accurately than any judge. But it does not experience mercy, remorse, suffering or redemption. It can recognize those concepts, describe them and perhaps model how human beings have historically responded to them. That is not the same as possessing them.
The distinction is central to the argument I have made elsewhere about AI and criminal justice: calculation is not judgment. A sentencing guideline calculates; a judge sentences. A forensic instrument measures; an expert interprets. A risk model predicts; a prosecutor exercises discretion. An algorithm correlates; a jury decides.
The machine can assist the judgment. It should not become the institution that judges.
Human Imperfection May Actually Be Part of the Safeguard
This leads to a paradox that is easy to miss. The imperfections of human beings are simultaneously an argument for using artificial intelligence and an argument against allowing artificial intelligence to replace us.
We need AI precisely because human beings make mistakes. Let the machine challenge the eyewitness. Let it discover that the witness could not possibly have seen what she claims to have seen. Let it identify inconsistencies in police reports, test a prosecutor’s timeline, compare DNA probabilities, analyze thousands of hours of video and find exculpatory evidence buried inside millions of pages of discovery.
Perhaps AI should even be permitted to tell the jury what conclusion it reaches and why, provided the system satisfies rigorous standards of reliability, transparency and adversarial testing. But its conclusion should remain evidence rather than a verdict.
Human beings must remain free to reject it because justice requires more than optimization. It requires someone who can be held responsible for the decision and who understands that the subject of that decision is not a datapoint but another human being.
The legal system has never been built on the assumption that people are infallible. Quite the opposite. Witnesses take oaths because people can lie. Cross-examination exists because witnesses can be mistaken. Appeals exist because judges can be wrong. Juries deliberate because reasonable people can disagree.
Human fallibility is not an undiscovered bug in the justice system. The system was designed around it.
The Most Dangerous AI May Be the One That Is Almost Always Right
The greatest danger may therefore arrive not when artificial intelligence performs badly but when it performs extraordinarily well. If an AI system is wrong half the time, no judge will trust it. If it is right 99.9 percent of the time, disagreeing with it begins to look irrational. Human oversight can then become ceremonial: the machine reaches the conclusion, and the human being simply signs underneath it.
That is when the lesson of Minority Report becomes most important. The minority report matters precisely because the majority is usually right. Dissent matters because consensus can become complacency. Reasonable doubt matters because an individual defendant should never disappear inside the statistical reliability of a system. A society genuinely committed to human judgment must therefore preserve something AI may never understand: the right to be the exception.
Keep the Machine. Keep the Minority Report. Keep the Human Being.
Artificial intelligence could become one of the greatest tools for justice ever developed. We should use it aggressively to expose weaknesses in evidence, identify wrongful convictions, test eyewitness accounts, analyze forensic evidence and challenge assumptions that human investigators have failed to question. But we should never confuse superior calculation with the authority to condemn another human being.
That distinction becomes more important, not less important, as AI improves. The real test will come when the machine is more accurate than we are—when every statistical measure says guilty, every algorithm agrees and the human being sitting in the jury box nevertheless says, “I am not convinced.”
Minority Report imagined a society that regarded that dissenting conclusion as a flaw in the system. Our legal tradition should recognize it as something very different. It may be the system working exactly as intended.
Artificial intelligence can give us the majority report. It can calculate probabilities beyond the capacity of any human mind, expose contradictions we would never notice and perhaps prevent some of the terrible errors human beings have made for centuries. But when the government proposes to take away another person’s liberty—or perhaps someday even life—someone who possesses conscience, accountability, compassion and the capacity for doubt must remain responsible for the decision.
The machine may ultimately be right more often than we are. That still does not make it human. And justice, however imperfectly administered, must remain a human responsibility.
Frequently Asked Questions About AI and Human Judgment in Criminal Justice
What question does this article explore?
The article asks whether society should allow AI to determine guilt if it eventually becomes more accurate than human judges or juries. Peter argues that accuracy alone does not settle who should hold the authority and responsibility to judge another person.
How does Minority Report connect to the article’s argument?
Peter focuses on the film’s treatment of a dissenting prediction as an inconvenience to an apparently reliable system. He uses that premise to explore why disagreement and doubt must remain meaningful even when a machine’s conclusions appear persuasive.
Are FORECAST and United States v. Mercer real?
No. The article expressly presents FORECAST and United States v. Mercer as elements of a fictional future case. The hypothetical accuracy figures illustrate an ethical dilemma; they are not reported performance results for an existing AI system.
Does Peter argue against using AI in criminal justice?
No. He describes potential uses that include challenging eyewitness accounts, identifying inconsistencies, analyzing forensic evidence and finding exculpatory material. His distinction is between AI assisting human judgment and AI becoming the institution that delivers the verdict.
Why does the article warn about apparent mathematical neutrality?
Peter argues that models trained on historical criminal justice information may reproduce biases embedded in that information. A precise numerical score can make those inherited assumptions appear more objective than they are.
Why might a highly accurate system still need meaningful human oversight?
The article argues that strong overall performance can make people reluctant to challenge an individual result. Peter’s concern is that oversight could become a formality unless people retain the ability to question the system, reject its conclusion and accept responsibility for the decision.
